By Benjamin Guggenheim

Not a subscriber? Sign up here to get this newsletter in your inbox.

• Exclusive: The Trump administration and the AI industry have been discussing a capability framework for U.S. open-source models based on the current capabilities of leading Chinese open-source models.
• The AI industry expects Chinese Mythos-class models to be available for free download on the internet in six to 12 months. That greatly complicates the future of open-source models in the U.S.
• Energy consumption by data centers may be overestimated in the long run. Companies are experimenting with specialized conductors and light waves to transmit information in a much more energy-efficient way.

This is Benjamin Guggenheim. Welcome to WP Intelligence’s AI & Tech Brief, where we examine the transformative technology of artificial intelligence at the intersection of innovation, policy and power.

I’m back from a rejuvenating vacation in Portugal and Norway, where I refrained from looking at any AI news whatsoever and deleted the X app from my phone. Help me get up to speed at Benjamin.Guggenheim@washpost.com or on Signal at ben_guggenheim.88.

The Lead Brief

Exclusive: The administration and industry groups have been discussing a capability framework for U.S. open-source models based on the current capabilities of leading open-source Chinese models, a person close to the discussions tells me.

The proposal would streamline U.S.-made models to market (both open-source and “closed,” licensed models) if their capabilities are equal to or below the capabilities of leading open-source Chinese models, the person, who spoke on the condition of anonymity to share details of the private conversations, said.

Currently, the conversations are not centered around a separate, forthcoming executive order on open-source AI models (as speculation online has suggested), but rather clarifying guidance on the AI executive order that the Trump administration already issued in June.

The national security issue: Open-source models, or AI systems whose architectures are fully published online, are poised to become much more consequential from a policy perspective in the coming months — and there are competing pressures that could spur the adoption of open models or squelch them in their adolescence.

For example, lawmakers have become increasingly concerned with the use of Chinese open-source models by prominent American companies such as Airbnb.

ADVERTISEMENT

According to Daniel Remler, a senior fellow at the Center for a New American Security, the concern is not only that Chinese open-source models could act as “Trojan horses” for malicious software, but also that developers have intentionally left back doors in models that could be exploited by the Chinese Communist Party.

“These models have documented serious security vulnerabilities that make them more susceptible to adversarial attack …. These are vulnerabilities that Chinese security services could exploit for passing intelligence cyberattacks,” said Remler, who added that there has been some research into “backdoor” sleeper agents in Chinese AI models.

These national security concerns have prompted companies, such as Reflection AI, to argue that the U.S. needs a leading open-source model that can act as an alternative to the leading Chinese closed models.

The basic value proposition of the Chinese models is that they are cheaper and easier to access and modify. Businesses across the globe should have a U.S. open-source option to build off of without worrying about adversarial compromise, open-source advocates say.

The page for the smartphone app DeepSeek is seen on a smartphone screen in Beijing. (AP Photo/Andy Wong)

ADVERTISEMENT

The open-source conundrum

The Trump administration’s moves to pressure Anthropic into retracting its leading model, Fable, indicates that the government is nowhere close to developing a tried-and-true method for vetting frontier AI models before their release.

Open-source models are particularly tricky in this regard. Unlike closed models — which the government can force companies to retract if necessary — the publication of open-source models, by definition, means that they are out in the world for good.

The administration won’t get a second chance at patching vulnerabilities or preventing jailbreaks once an open-source model is out there.

Helen Toner, the executive director at Georgetown’s Center for Security and Emerging Technology and a former OpenAI board member, also points out that companies can’t monitor the use of open-source models the way that companies can with closed models.

“With open source, they can try and sort of train safeguards into the model, but it’s very easy to train them out again. So I do think that should affect the risk calculus,“ said Toner.

Given that policymakers predict that Chinese models are approximately six months behind the leading models developed by Anthropic and OpenAI, some experts, such as machine learning engineer Nathan Lambert, predict that the administration will soon permanently ban advanced open-source models to mitigate cyber and bio threats.

“The most likely incoming action is to ban or indefinitely delay any open-weights model meaningfully above the capability level in the range of GPT 5.5, Claude Opus 4.8, or GLM-5.2,” wrote Lambert Sunday in his newsletter “Interconnects,” which is widely read by engineers at the frontier labs.

“With the consistent capability gap, this should be within the next 6 months,” Lambert writes.

Anthropic CEO Dario Amodei has also been sounding the alarm about persistent “distillation” attacks where Chinese companies attempt to copy the capabilities of leading American AI models. Amodei has said that it’s also a “serious concern” that open-source models with Mythos-class cyber capabilities are coming down the pike.

“What I do worry about with some of these laggard models is the risks of them, where we have Mythos-class cyber capabilities,” Amodei said in an interview on Bloomberg in June. “Months from now, Mythos-class cyber capabilities may just be available for anyone to download.”

The upshot: A capability framework that would streamline the release of U.S. open-source models would likely require the Commerce Department’s Center for AI Standards and Innovation to develop a procedure for comparing U.S. open models to Chinese ones.

However, a streamlined process doesn’t change the fact that the clock is ticking on open-source models that can pose serious cyber and biological risks.

“I don’t think there’s a straightforward line from, ‘Open-source models will reach Mythos-level capabilities,’ and then they will be permanently banned,” said Toner. “The online security environment will have changed also, and we will have had several months to kind of get used to having this level of of capability.”

Open-source advocates will argue that, if such Chinese models exist, then open-source U.S. models should, too. But it’s not out of the question that the administration or Congress will eventually move to ban Chinese open-source models entirely.

“I think it’s probably going to be difficult to actually ban these Chinese models, but that’s not to say Congress won’t try,” said a senior GOP aide who works on AI policy, and spoke on the condition of anonymity to speak candidly.

Data center energy forecast

Data centers may end up needing less electricity than the AI industry forecasts. That’s because innovators are starting to question key assumptions that have undergirded the operation of data centers for years.

As WP Intelligence Lead Energy Analyst Kathryn Clay outlines in a new report, those assumptions are threefold:

1. That electricity must flow through copper.
2. That data must move electrically.
3. That most AI needs to run on cloud data centers.

The traditional architecture: Cheap and malleable copper has been the obvious choice for electrical components for decades, but soaring costs of compute are prompting companies to look at more energy-efficient conductors.

One option includes high-temperature superconductors (HTS), which operate at near-zero degrees and practically eliminate all electrical friction. The cost of the specialized refrigeration systems needed to run HTS has historically been prohibitive, but the unprecedented demands for AI power are changing the system economics.

Transmitting data: Another option is to forgo transmitting electrical signals altogether and use “photonics,” which transmits information through pulses of light. (Photonics is also one of the most promising architectures for quantum computing.)

Several photonics developers for data centers, such as the Mountain View, California-based Lightmatter, argue that photonics save computing resources because the light beams transmit information much more quickly than material conductors. Photonics also reduces the cooling requirements for traditional semiconductors.

Where AI is operated: Start-ups and major companies such as Apple, Qualcomm and Nvidia are experimenting with using “latent compute” on local devices to run AI rather than transmitting AI requests to a hyperscale data center.

The outlook: Utilities warned in the 1980s that controlling acid rain would require extraordinarily expensive pollution controls. But costs ended up falling dramatically by the 2000s because of innovations in emission-control technologies and emissions-trading schemes.

Similarly, technological breakthroughs in fiber optic cables meant that surging internet traffic (at the dawn of the internet) required far less communications infrastructure buildout than originally anticipated.

Kathryn points out that when markets face large technical constraints, innovators tend to produce solutions that forecasts fail to anticipate. Kathryn recommends that companies stress-test long-term capital investments against multiple technology scenarios.

This newsletter is published by WP Intelligence, The Washington Post’s subscription service for professionals that provides business, policy and thought leaders with actionable insights. WP Intelligence operates independently from The Washington Post newsroom. Learn more about WP Intelligence.